In the realm of medical imaging, a trio of researchers from the Bangladesh University of Engineering and Technology (BUET) have developed a novel deep learning framework that promises to enhance the consistency and accuracy of ultrasound strain elastography (USE). Shourov Joarder, Tushar Talukder Showrav, and Md. Kamrul Hasan have introduced MUSSE-Net, a multi-stage, unsupervised deep learning approach designed to overcome the challenges that have limited the clinical application of USE.
Ultrasound strain elastography is a non-invasive imaging technique used to assess the mechanical properties of tissues, offering valuable diagnostic insights across various clinical applications. However, its widespread use has been hindered by issues such as tissue decorrelation noise, a lack of ground truth data, and inconsistent strain estimation under different deformation conditions. The researchers’ proposed solution, MUSSE-Net, aims to address these barriers through a sophisticated, sequential deep learning framework.
At the heart of MUSSE-Net lies USSE-Net, an end-to-end multi-stream encoder-decoder architecture that processes pre- and post-deformation radiofrequency (RF) sequences in parallel to estimate displacement fields and axial strains. The architecture incorporates several innovative components, including a Context-Aware Complementary Feature Fusion (CACFF)-based encoder, a Tri-Cross Attention (TCA) bottleneck, and a Cross-Attentive Fusion (CAF)-based sequential decoder. To ensure temporal coherence and strain stability across varying deformation levels, the architecture employs a tailored consistency loss.
Following the initial strain estimation, MUSSE-Net includes a secondary residual refinement stage that further enhances accuracy and suppresses noise. The researchers validated their framework extensively using simulation, in vivo, and private clinical datasets from the BUET medical center. The results demonstrated that MUSSE-Net outperformed existing unsupervised approaches, achieving state-of-the-art performance metrics. Notably, on the BUET dataset, MUSSE-Net produced strain maps with enhanced lesion-to-background contrast and significant noise suppression, yielding clinically interpretable strain patterns.
While this research is primarily focused on medical imaging, the advanced deep learning techniques and architectures developed by the BUET researchers could potentially find applications in the energy sector. For instance, similar approaches could be adapted for improving the accuracy and consistency of data analysis in energy exploration, where the interpretation of seismic or other subsurface imaging data is crucial. Additionally, the noise suppression capabilities of MUSSE-Net could be leveraged to enhance the quality of data collected from energy infrastructure, such as pipelines or wind turbines, where sensor data can often be affected by noise and other artifacts.
The research was published in the IEEE Transactions on Medical Imaging, a prestigious journal dedicated to the publication of original contributions in the field of medical imaging. The full paper can be accessed on the IEEE Xplore digital library.
This article is based on research available at arXiv.